daTechGuy · Oct 4, 2026 · Oct 3, 2026 · Oct 3, 2026
@@ -12,6 +12,7 @@ npm run build          # tsc -b + vite build  (tsc only type-checks src/, NOT te
 npm run lint           # ESLint over everything (7 known react-hooks/exhaustive-deps warnings, all pre-existing upstream)
 npm test               # vitest (happy-dom). CI runs `test:coverage`
 npm run validate:imscc -- file.imscc   # offline structural check of a Canvas cartridge
+GEMINI_API_KEY=... npm run smoke:gemini # live check of the Gemini backend (key never printed)
 
 # Canvas-importable course from a topic + a Canvas template, headless
 npx tsx scripts/canvas-course.ts --topic "..." --template ./template.imscc --chapters 8 --output ./out
@@ -56,7 +57,7 @@ src/
 ├── types/                      # course.ts (all data interfaces), template.ts, outline.ts, generation.ts
 ├── store/                      # Zustand
 │   ├── courseStore.ts          #   course state (persisted to IndexedDB via idbStorage)
-│   ├── apiStore.ts             #   API keys, provider, research backend, advancedMode (persisted)
+│   ├── apiStore.ts             #   API keys, provider, research backend, advancedMode (persisted); selectActiveLlm()
 │   ├── templateStore.ts        #   uploaded Canvas templates (parsed + raw .imscc blob)
 │   └── uiStore.ts              #   transient UI/generation state, in-flight map
 ├── pages/                      # Landing, Setup, Syllabus, Research, Build, Export, TemplatePreview
@@ -66,7 +67,7 @@ src/
 │   ├── setup/, syllabus/, export/, layout/, shared/
 ├── prompts/                    # prompt builders (syllabus, chapter, quizzes, slides, templateChapter…)
 ├── services/
-│   ├── llm/                    #   provider router: anthropic.ts, ollama.ts, index.ts (the real streaming layer)
+│   ├── llm/                    #   provider router: anthropic.ts, gemini.ts, ollama.ts, index.ts (the real streaming layer)
 │   ├── claude/                 #   SDK client, model ids, thinking budgets; streaming.ts is a thin facade over llm/
 │   ├── research/               #   backends: anthropic (web search), tavily, wikipedia; runResearch() dispatch
 │   ├── academic/               #   Crossref / Semantic Scholar / Unpaywall source enrichment
@@ -85,7 +86,9 @@ tests/                          # vitest; fixtures/ holds real Canvas-made .imsc
 
 - `tsconfig.app.json` has `erasableSyntaxOnly: true` — no parameter properties; use class field declarations. `noUnusedLocals` / `noUnusedParameters` are on — always run `npm run build`.
 - **`tsc -b` only checks `src/`.** Test files are transpiled by vitest but not type-checked; keep them type-correct anyway.
-- **Providers (Claude / Ollama Cloud) go through `src/services/llm/`.** `apiStore.provider` picks the backend. Call sites pass Anthropic model ids (`MODELS.opus` …); `resolveProvider` **ignores any `claude-*` model id on the Ollama path** and uses `apiStore.ollamaModel`. Never forward a Claude model id to Ollama.
+- **Providers (Claude / Gemini / Ollama Cloud) go through `src/services/llm/`.** `apiStore.provider` picks the backend. Call sites pass Anthropic model ids (`MODELS.opus` …) and `claudeApiKey`; `resolveProvider` **ignores any `claude-*` model id on the Gemini/Ollama paths** and swaps in that provider's own key and model. Never forward a Claude model id to another provider.
+- **Never gate on `claudeApiKey`.** "Can we generate?" is `selectActiveLlm(useApiStore()).hasKey` (also `.label` / `.keyValid`). Gating on the Claude key silently blocked Ollama/Gemini users (the syllabus never auto-started). Adding a provider = a backend in `services/llm/`, fields in `apiStore`, a pill + key in the Setup keys dialog, a case in `selectActiveLlm` and `resolveProvider`, CLI flags, and tests.
+- **Gemini** (`services/llm/gemini.ts`) is called directly from the browser — Google answers CORS preflights for the `x-goog-api-key` header, so no proxy (unlike Ollama). Key goes in a header, never the URL. It ignores Anthropic server tools, so "Claude web search" research still needs a Claude key; Tavily/Wikipedia work. Thought summaries are requested only with a thinking budget and retried once without on a 400 that mentions thinking. Model ids change often — the model is a free-text setting. Unit tests mock `fetch`; `npm run smoke:gemini` (key from `GEMINI_API_KEY`) checks the live API.
 - **Opus 4.8** uses adaptive thinking (`thinking: {type:'adaptive'}` + `output_config.effort`); manual `budget_tokens` is rejected on Opus ≥4.7. Sonnet/Haiku still use manual budgets. Pass `signal` for cancellation and don't retry an aborted call.
 - **`advancedMode`** (apiStore) is "Canvas-focused" (Reading / Practice / Quizzes / Discussion) vs "Everything" (+ Challenge / Activities / Audio / Slides). It gates the visible tabs **and** batch / "Generate all" generation **and** the sidebar's N-of-M count — keep them in step.
 - **Don't put an inline `onClose` in an effect dependency list.** `CodexModal` once re-focused the dialog on every keystroke because its effect depended on the caller's inline arrow; it now holds `onClose` in a ref. Same trap anywhere an effect side-effect (focus, scroll) follows a changing callback.

@@ -26,7 +26,7 @@ A fork of [ClassBuild](https://github.com/jtangen/classbuild) by Jason Tangen, r
 | Batch generation: "Generate All Canvas Modules" for the whole course at once | Build header |
 | Export an `.imscc` that mirrors the template's structure: verbatim modules carry through untouched, pattern modules replaced by generated content, LTI links + web_resources passed through | "Export for Canvas (.imscc)" button |
 | Upload a course-outline `.docx` and have an LLM extract title / description / course information / course materials → populates Canvas's Syllabus tab body at export time | Setup page |
-| Pluggable LLM provider — **Anthropic (Claude)** or **Ollama Cloud** for course-content generation | Setup → Course-content provider |
+| Pluggable LLM provider — **Anthropic (Claude)**, **Google Gemini** or **Ollama Cloud** for course-content generation | Setup → Course-content provider |
 | Pluggable research backend — **Claude web search**, **Tavily**, or **Wikipedia** | Setup → Research backend |
 | Advanced-mode toggle — hides multimedia outputs (slides, audio, infographic, weekly challenge, activities) by default to keep the Canvas-focused happy path tight | Setup page bottom |
 
@@ -48,6 +48,7 @@ Open [localhost:5173](http://localhost:5173).
 | Provider | Required when… | What it does |
 |---|---|---|
 | Anthropic Claude | LLM provider = Anthropic, **OR** research backend = Claude web search | Course-content generation; Claude's built-in web search for the Research stage |
+| Google Gemini | LLM provider = Gemini | Course-content generation, called straight from the browser (Google allows CORS — no proxy). Get a key at aistudio.google.com/apikey. The model id is editable (default `gemini-3.8-flash`). Doesn't run Claude's web-search tool, so pair it with Tavily or Wikipedia research. On Google's free tier prompts may be used to improve its products — check your plan |
 | Ollama Cloud | LLM provider = Ollama Cloud | Course-content generation. Free tier doesn't include cloud models — see ollama.com/settings/keys |
 | Tavily | Research backend = Tavily | Web search for the Research stage. Free tier covers ~1,000 searches/month |
 | OpenAI | Advanced mode + you want slide images / chapter figures | Image generation (gpt-image-2, optional) |
@@ -154,12 +155,21 @@ Flags:
 | `--length` | `standard` | `concise` / `standard` / `comprehensive` |
 | `--cohort` | `60` | Class size |
 | `--notes` | — | Additional learner context |
-| `--provider` | `anthropic` | LLM provider — `anthropic` or `ollama`. With `ollama`, set `OLLAMA_API_KEY`; with `anthropic`, set `ANTHROPIC_API_KEY` |
+| `--provider` | `anthropic` | LLM provider — `anthropic`, `gemini` or `ollama`. Set `ANTHROPIC_API_KEY`, `GEMINI_API_KEY` or `OLLAMA_API_KEY` to match |
 | `--ollama-model` | `gpt-oss:120b-cloud` | Ollama model used when `--provider ollama` |
+| `--gemini-model` | `gemini-3.8-flash` | Gemini model used when `--provider gemini` |
 | `--concurrency` | `3` | Parallel chapters during Canvas Module generation |
 | `--syllabus` | — | Path to existing `syllabus.json` (skip regeneration) |
 | `--skip-canvas-module` | `false` | Skip per-chapter content generation (for fast manifest-only test runs) |
 
+Gemini example (`GEMINI_API_KEY` must be set in your environment):
+
+```bash
+npx tsx scripts/canvas-course.ts \n  --provider gemini \n  --topic "Statistics for Data Science" \n  --template ./template.imscc \n  --chapters 8 \n  --output ./output/stats
+```
+
+Before relying on a Gemini model, run `npm run smoke:gemini` (reads `GEMINI_API_KEY` from your environment): it exercises plain text, streaming, thinking, structured-JSON quiz output and the error paths against the live API.
+
 Ollama example:
 
 ```bash
@@ -172,7 +182,7 @@ OLLAMA_API_KEY=... npx tsx scripts/canvas-course.ts \
   --output ./output/stats
 ```
 
-If both `OLLAMA_API_KEY` and `ANTHROPIC_API_KEY` are set, outline-DOCX extraction always uses Claude (its instruction-following on the small JSON-extraction task is more reliable) regardless of `--provider`.
+If `ANTHROPIC_API_KEY` is set alongside `--provider gemini` or `ollama`, outline-DOCX extraction always uses Claude (its instruction-following on the small JSON-extraction task is more reliable) regardless of `--provider`.
 
 ### Original ClassBuild CLI ([`scripts/generate-course.ts`](scripts/generate-course.ts))
 
@@ -195,6 +205,7 @@ src/
     llm/
       anthropic.ts           # Anthropic SDK streaming
       ollama.ts              # Ollama Cloud NDJSON streaming
+      gemini.ts              # Google Gemini SSE streaming (direct from the browser)
   store/
     templateStore.ts         # uploaded templates (parsed + raw .imscc Blob)
   pages/
@@ -209,7 +220,7 @@ src/
 
 ## Built with
 
-React 19 · Vite 7 · TypeScript 5.9 · Tailwind CSS 4 · Zustand · JSZip · mammoth.js · Claude (Opus 4.8 / Sonnet 4.6 / Haiku 4.5) · Ollama Cloud · Tavily · OpenAI gpt-image-2 · ElevenLabs
+React 19 · Vite 7 · TypeScript 5.9 · Tailwind CSS 4 · Zustand · JSZip · mammoth.js · Claude (Opus 4.8 / Sonnet 4.6 / Haiku 4.5) · Google Gemini · Ollama Cloud · Tavily · OpenAI gpt-image-2 · ElevenLabs
 
 ## Contributing
 

@@ -34,7 +34,8 @@
     "test": "vitest run",
     "test:watch": "vitest",
     "test:coverage": "vitest run --coverage",
-    "validate:imscc": "tsx tools/validate-imscc.ts"
+    "validate:imscc": "tsx tools/validate-imscc.ts",
+    "smoke:gemini": "tsx tools/smoke-gemini.ts"
   },
   "dependencies": {
     "@anthropic-ai/sdk": "^0.74.0",

@@ -63,8 +63,9 @@ const { values } = parseArgs({
     concurrency: { type: 'string', default: '3' },
     syllabus: { type: 'string' }, // path to existing syllabus.json (skip regeneration)
     'skip-canvas-module': { type: 'boolean', default: false },
-    provider: { type: 'string', default: 'anthropic' }, // anthropic | ollama
+    provider: { type: 'string', default: 'anthropic' }, // anthropic | ollama | gemini
     'ollama-model': { type: 'string', default: 'gpt-oss:120b-cloud' },
+    'gemini-model': { type: 'string', default: 'gemini-3.8-flash' },
   },
   strict: true,
 });
@@ -80,13 +81,15 @@ if (!values.template) {
   process.exit(1);
 }
 
-type Provider = 'anthropic' | 'ollama';
+type Provider = 'anthropic' | 'ollama' | 'gemini';
 const PROVIDER: Provider =
-  values.provider === 'ollama' ? 'ollama' : 'anthropic';
+  values.provider === 'ollama' ? 'ollama' : values.provider === 'gemini' ? 'gemini' : 'anthropic';
 const OLLAMA_MODEL = values['ollama-model']!;
+const GEMINI_MODEL = values['gemini-model']!;
 
 const ANTHROPIC_API_KEY = process.env.ANTHROPIC_API_KEY;
 const OLLAMA_API_KEY = process.env.OLLAMA_API_KEY;
+const GEMINI_API_KEY = process.env.GEMINI_API_KEY;
 
 if (PROVIDER === 'anthropic' && !ANTHROPIC_API_KEY) {
   console.error('Error: ANTHROPIC_API_KEY environment variable is required for --provider anthropic');
@@ -96,12 +99,28 @@ if (PROVIDER === 'ollama' && !OLLAMA_API_KEY) {
   console.error('Error: OLLAMA_API_KEY environment variable is required for --provider ollama');
   process.exit(1);
 }
+if (PROVIDER === 'gemini' && !GEMINI_API_KEY) {
+  console.error('Error: GEMINI_API_KEY environment variable is required for --provider gemini');
+  process.exit(1);
+}
+
+/** Per-call provider overrides for the LLM router (the CLI never touches the browser store). */
+function providerOptions(p: Provider) {
+  return {
+    provider: p,
+    ollamaApiKey: p === 'ollama' ? OLLAMA_API_KEY : undefined,
+    ollamaModel: p === 'ollama' ? OLLAMA_MODEL : undefined,
+    geminiApiKey: p === 'gemini' ? GEMINI_API_KEY : undefined,
+    geminiModel: p === 'gemini' ? GEMINI_MODEL : undefined,
+  };
+}
 
 // The outline-DOCX extraction step always runs on Claude when available
 // (it's a one-shot JSON extraction that benefits from instruction-following
-// quality). If Anthropic isn't configured, fall back to Ollama.
-const OUTLINE_KEY = ANTHROPIC_API_KEY ?? OLLAMA_API_KEY!;
-const OUTLINE_PROVIDER: Provider = ANTHROPIC_API_KEY ? 'anthropic' : 'ollama';
+// quality). If Anthropic isn't configured, fall back to the chosen provider
+// (its key was validated above).
+const OUTLINE_PROVIDER: Provider = ANTHROPIC_API_KEY ? 'anthropic' : PROVIDER;
+const OUTLINE_KEY = ANTHROPIC_API_KEY ?? (PROVIDER === 'gemini' ? GEMINI_API_KEY! : OLLAMA_API_KEY!);
 
 const setup: CourseSetup = {
   topic: values.topic,
@@ -136,11 +155,11 @@ function printUsage() {
       '    [--level advanced-undergrad] \\\n' +
       '    [--length concise|standard|comprehensive] \\\n' +
       '    [--notes "Additional learner context"] \\\n' +
-      '    [--provider anthropic|ollama] \\\n' +
-      '    [--ollama-model gpt-oss:120b-cloud] \\\n' +
+      '    [--provider anthropic|ollama|gemini] \\\n' +
+      '    [--ollama-model gpt-oss:120b-cloud] [--gemini-model gemini-3.8-flash] \\\n' +
       '    [--concurrency 3] \\\n' +
       '    [--syllabus ./existing-syllabus.json]\n' +
-      '\nFor Ollama: set OLLAMA_API_KEY (and optionally ANTHROPIC_API_KEY for\noutline extraction quality if you have it).',
+      '\nFor Ollama: set OLLAMA_API_KEY; for Gemini: set GEMINI_API_KEY (optionally also ANTHROPIC_API_KEY for\noutline extraction quality if you have it).',
   );
 }
 
@@ -218,9 +237,7 @@ async function main() {
     const result = await extractOutlineFields({
       apiKey: OUTLINE_KEY,
       rawText,
-      provider: OUTLINE_PROVIDER,
-      ollamaApiKey: OUTLINE_PROVIDER === 'ollama' ? OLLAMA_API_KEY ?? undefined : undefined,
-      ollamaModel: OUTLINE_PROVIDER === 'ollama' ? OLLAMA_MODEL : undefined,
+      ...providerOptions(OUTLINE_PROVIDER),
     });
     outlineFields = result.fields;
     const found = Object.keys(outlineFields).filter(
@@ -248,9 +265,7 @@ async function main() {
         messages: [{ role: 'user', content: userMessage }],
         thinkingBudget: 'high',
         maxTokens: 16000,
-        provider: PROVIDER,
-        ollamaApiKey: PROVIDER === 'ollama' ? OLLAMA_API_KEY! : undefined,
-        ollamaModel: PROVIDER === 'ollama' ? OLLAMA_MODEL : undefined,
+        ...providerOptions(PROVIDER),
       },
       {},
     );
@@ -282,9 +297,7 @@ async function main() {
           courseTitle: syllabus.courseTitle,
           courseOverview: syllabus.courseOverview,
           examplePatternContent: template.examplePatternContent,
-          provider: PROVIDER,
-          ollamaApiKey: PROVIDER === 'ollama' ? OLLAMA_API_KEY! : undefined,
-          ollamaModel: PROVIDER === 'ollama' ? OLLAMA_MODEL : undefined,
+          ...providerOptions(PROVIDER),
         });
         const target = chapters.find((c) => c.number === ch.number);
         if (target) target.templateContent = content;

@@ -1,6 +1,6 @@
 import { useRef, useState } from 'react';
 import { useCourseStore } from '../../store/courseStore';
-import { useApiStore } from '../../store/apiStore';
+import { useApiStore, selectActiveLlm } from '../../store/apiStore';
 import { docxToText, extractOutlineFields } from '../../services/template/parseOutlineDocx';
 import { Button } from '../shared/Button';
 import { friendlyError } from '../../utils/errors';
@@ -22,25 +22,23 @@ const FIELD_HINTS: Record<keyof OutlineFields, string> = {
 
 export function CourseOutlineUpload() {
   const { outlineFields, outlineRawText, setOutlineFields, setOutlineRawText } = useCourseStore();
-  const { provider, claudeApiKey, ollamaApiKey } = useApiStore();
+  const llm = selectActiveLlm(useApiStore());
   const [busy, setBusy] = useState(false);
   const [error, setError] = useState<string | null>(null);
   const [open, setOpen] = useState(false);
   const fileInputRef = useRef<HTMLInputElement>(null);
 
-  const llmKey = provider === 'ollama' ? ollamaApiKey : claudeApiKey;
-
   const handleFile = async (file: File) => {
-    if (!llmKey.trim()) {
-      setError('Add an API key (Claude or Ollama) before uploading — the outline is parsed by the active LLM provider.');
+    if (!llm.hasKey) {
+      setError(`Add your ${llm.label} API key before uploading — the outline is parsed by the active LLM provider.`);
       return;
     }
     setBusy(true);
     setError(null);
     try {
       const rawText = await docxToText(file);
       setOutlineRawText(rawText);
-      const result = await extractOutlineFields({ apiKey: llmKey, rawText });
+      const result = await extractOutlineFields({ apiKey: llm.apiKey, rawText });
       setOutlineFields(result.fields);
       setOpen(true);
     } catch (err) {

@@ -1,7 +1,7 @@
 import { useState, useCallback, useMemo, useRef } from 'react';
 import { motion } from 'framer-motion';
 import { useCourseStore } from '../store/courseStore';
-import { useApiStore } from '../store/apiStore';
+import { useApiStore, selectActiveLlm } from '../store/apiStore';
 import { useUiStore } from '../store/uiStore';
 import { streamMessage } from '../services/claude/streaming';
 import { MODELS } from '../services/claude/client';
@@ -89,7 +89,9 @@
 export function ExportPage() {
   const { syllabus, chapters, researchDossiers, addChapter, updateChapter, setSlideImage, setup, updateSetup, curriculumMap, setCurriculumMap, outlineFields } = useCourseStore();
   const currentThemeId = (setup.themeId as ChapterThemeId | undefined) ?? DEFAULT_CHAPTER_THEME_ID;
-  const { claudeApiKey, openaiApiKey } = useApiStore();
+  const apiState = useApiStore();
+  const { claudeApiKey, openaiApiKey } = apiState;
+  const llm = selectActiveLlm(apiState);
   const { isGenerating, setIsGenerating, setError, error } = useUiStore();
   const [generatingChapter, setGeneratingChapter] = useState<number | null>(null);
   const [isPublishing, setIsPublishing] = useState(false);
@@ -121,7 +123,7 @@
    } catch {
      downloadFile(chapter.practiceQuizData, `quiz-${chapterNum}-${sanitizeFilename(chapter.title)}.txt`, 'text/plain');
    }
  }, [chapters, syllabus]);

  const handleDownloadSlides = useCallback(async (chapterNum: number) => {
    const chapter = chapters.find(c => c.number === chapterNum);
@@ -524,7 +526,7 @@
     curriculumMap.syllabusHash !== currentSyllabusHash;
 
   const generateLearningOutcomes = useCallback(async () => {
-    if (!syllabus || !claudeApiKey || isGeneratingOutcomes) return;
+    if (!syllabus || !llm.hasKey || isGeneratingOutcomes) return;
     setIsGeneratingOutcomes(true);
     setOutcomesError(null);
     outcomesStreamRef.current = '';
@@ -564,6 +566,7 @@
   }, [
     syllabus,
     claudeApiKey,
+    llm.hasKey,
     isGeneratingOutcomes,
     currentSyllabusHash,
     setCurriculumMap,
@@ -666,7 +669,7 @@
      setIsGenerating(false);
      setGeneratingChapter(null);
    }
  }, [syllabus, chapters, claudeApiKey, researchDossiers, setup.chapterLength, addChapter, updateChapter, isGenerating, setIsGenerating, setError]);

  // --- Per-class bundle ---
  //
@@ -1011,7 +1014,7 @@

    const blob = await zip.generateAsync({ type: 'blob' });
    saveAs(blob, `${courseName}.zip`);
  }, [syllabus, chapters, setup.themeId, buildCurriculumMapCsv, curriculumMap]);

  const handlePublish = useCallback(async () => {
    if (!syllabus || chapters.length === 0) return;
@@ -1406,7 +1409,7 @@
           onGenerate={() => void generateLearningOutcomes()}
           onDownloadHtml={handleDownloadOutcomesHtml}
           onDownloadCsv={handleDownloadOutcomesCsv}
-          canGenerate={!!claudeApiKey && !isGenerating}
+          canGenerate={llm.hasKey && !isGenerating}
         />
 
         <div className="cb-folio">— 05 —</div>